Originally published 8 May 2008; corrected 9 June 2008
www.sciencemag.org/cgi/content/full/1154456/DC1
Supporting Online Material for
Predictive Behavior Within Microbial Genetic Networks
Ilias Tagkopoulos, Yir-Chung Liu, Saeed Tavazoie*
*To whom correspondence should be addressed. E-mail: [email protected]
Published 8 May 2008 on Science Express
DOI: 10.1126/science.1154456
This PDF file includes:
SOM Text
Figs. S1 to S21
Table S1
References
Correction (9 June 2008): Minor corrections to data (in SOM text and in legend of fig. S17).
1. Evolution in Variable Environments (EVE)
1.1 Simulation environment
In our simulation framework (EVE: Evolution in Variable Environments) populations of in silico
organisms compete and evolve under user-specified dynamic environments. Each organism (cell)
contains a biochemical network whose chemical species and their interactions/transformations are
modeled according to the “central dogma”. Simulations utilized multi-node supercomputer clusters
(BlueGene/L and Beowulf running on an average of 500 node workload for over 2 years).
Recent studies (S1-5) have made significant progress in simulating biochemical networks and extracting
novel insights in the context of focused biological questions (e.g. circadian rhythms, chemotaxis,
segmentation, etc.). Extending from these works, the EVE framework integrates many features that
improve the biochemical, evolutionary, and ecological realism of our simulations, features that are crucial
for simulating microbial regulatory networks in the context of interactions with the environment.
From the biochemical perspective, dynamics are modeled as stochastic events, not ordinary differential
equations, a feature that accounts for phenomena arising from small numbers of molecules. Additionally,
We take into account the full range of molecular species and their interactions, including genes, mRNAs,
siRNAs, miRNAs, proteins, and modified proteins. From the evolutionary perspective, EVE is designed to
simulate growth, mutation, and selection in a highly realistic fashion. EVE uses an asynchronous
simulation framework, where events such as cell-division, death, mutations etc., occur in a parallel,
asynchronous fashion. Another important distinction is the way mutations are modeled, having a range of
types and magnitudes that map to specific biological counterparts. Another important dimension of
realism is the capacity for gene-duplication and divergence which is a recurring theme in the evolution of
complex behavior in many of our experiments.
Another very important advantage of EVE—and essential for addressing our primary thesis here—is the
way in which the evolving organisms interact with their environments, sensing and responding to
dynamic and temporally structured signals. In this context, our organisms are not evolved to explicitly
implement specific functions (e.g. an oscillator), rather they evolve to maximize fitness through learning
the probabilistic relationships between signals they can sense and the abundance of resources. For ease
of interpretation, our simulations here have focused on sharply structured environments that correspond
to familiar logic gates (e.g. delayed dynamic XOR). However, real microbial habits are likely characterized
by probabilistic relationships amongst a large number of variables that show rather complex and fuzzy
relationships with fitness. This is a fundamental departure from the traditional view of the relationship
between regulatory-network structure and function. Our simulation framework provides the capacity to
explore this new paradigm.
The details of the EVE simulation environment are provided below.
1.2 Network components
Each organism comprises three types of nodes with distinct functionalities:
•
•
•
Gene/RNA Nodes
Protein Nodes
Modified Protein Nodes
Gene/RNA Nodes capture gene regulation and transcription of RNA (mRNA or siRNA). Protein Nodes
summarize the translation of mRNA to proteins and model post-transcriptional regulation. Modified
Protein Nodes capture the modification (phosphorylation, acetylation, etc) of proteins and other
conformational changes that may lead to gain/loss of function. The value of each node is a discrete
variable that corresponds to the number of node molecules present at a particular time. We define as
node triplet any set of a Gene/RNA Node and its corresponding Protein and Modified Protein Nodes.
1.3 Network Topology
The essence of each artificial organism is its internal interaction network. This can be depicted as a
directed graph where each vertex is a Gene/RNA, Protein or Modified Protein node, and each edge a
regulatory dependency between the two vertices that it connects. Edges are directional and their values
can be either negative or positive depending on the type of interaction. A quantitative representation of
the resulting graph is the interaction matrix W, where any non-zero value wij denotes the strength and
regulation type (negative – positive) of node i by node j. Figure S1 depicts a randomly created internal
network and its interaction matrix.
1.4 A probabilistic dynamical model
We created a probabilistic dynamical model of intracellular biochemical interactions that regards an
organism as a well-stirred and spatially homogeneous reaction chamber. This allows us to efficiently
capture stochastic phenomena, for example, noisy gene expression caused by a small number of
molecules. However, due to lack of spatial structure, this simulation framework is unsuitable when
compartmentalization and spatial localization is important to the function of the network/organism.
The general framework is similar to a dynamic Monte Carlo algorithm (S6) with a random selection
method where in each time unit the probability of each event (node creation, deletion, mutation, molecule
creation and degradation) is calculated and a particular event occurs if its probability is found to be
higher than a randomly generated number. Although this approach can be computationally more
intensive when compared to first reaction algorithms (S7), it requires less memory and is easier to
parallelize in multi-processor environments.
1.5 Expression of RNAs, proteins, and modified proteins
The value of each node corresponds to the total number of molecules of a particular type that are present
at that given time. The value of each node may fluctuate but always by one molecule at a time when
measured in consecutive time points. This fluctuation is a result of molecule creation, transformation or
degradation. To calculate the molecule production probability (i.e. the probability that a new molecule
will be created) we use two levels of sigmoid functions: One to capture the effect of each individual
regulator on the target node, and a second to compute the sum of those effects. More specifically to
calculate the molecule production probability of node i we first calculate the function:
, … , , , … , , , 1 · tanh ∑$ · , , ̃ "# Where the function , , ̃ " corresponds to the regulatory effect of node j on node i:
1
, , ̃ " · (1 tanh )
2
* +*
,+*
-.
&
In the previous equations, n is the total number of nodes, basali the basal expression parameter, /01 the
value of the ith row and jth column in the interaction matrix, 21 the value of node j, 30 and 40 the midpoint
and slope of the main target-specific sigmoid function, 301 and 5,67 the midpoint and slope of the regulatorspecific sigmoid function.
In the case of protein translation and modification, the molecule production probability also depends on
the number of substrate molecules (i.e. RNA, proteins) that are present in the environment. Additionally,
a realistic model has to capture the saturation effects of the translation and modification machinery (i.e.
ribosomes, enzymes) at high molecular concentrations. For this reason is multiplied by yet another
function 8 that is also a sigmoid in the case of protein translation and modification:
9 8 , · 8 Where
8 :
1
· ;1 tanh
<
2
1
+ =>?=+@?/
@B++C=+@?
Finally, the molecule production probability is given by:
+ =>?C>+D=+@?
+ 9 H 0
+ 0 I 9 I 1
+ 9 J 1
0
E F9
1
Figure S2 gives a general schematic of the expression model and depicts the effect various values of
midpoint and slope have on the sigmoid function +* * , +* , , +* ".
In our experiments, the parameters 3 and 4 of function 8 take the heuristic values 10 and 4
respectively. Although RNA molecules are not consumed during translation, each modification event
converts one unmodified protein molecule to its modified form and thus each protein modification causes
its (unmodified protein) pool to be reduced by one. Proteins that get modified have a low (<0.1)
probability of returning to their unmodified state. Although external regulation of the modified to
unmodified transitions is a feature in EVE, it was disabled in the experiments we present here.
1.6 Degradation
Degradation of molecules occurs in a somewhat similar manner. The probability that a molecule of node i
is degraded at any given time unit is proportional to the number of molecules present and is given by:
1 DKLMN
EKLMN , KLMN , KLMN , DKLMN " 2
1 DKLMN
2
· tanh )
KLMN
-
KLMN
Where 20 , OPQR0 , 3PQR0 , 4PQR0 are the number of molecules present, node specific degradation
parameter, degradation midpoint and slope respectively (
KLMN 50 ?B KLMN 4 in the
experiments presented here).
1.7 Triplet duplication and deletion
A Node triplet can be created or destroyed at any time unit during the course of an experiment. During
triplet duplication, a new triplet is created by cloning (duplicating) an existing one. The new triplet
initially has exactly the same associations and parameters as the old one, although it may eventually take
on its own course during evolution. Similarly, during a triplet deletion event a whole triplet is deleted. It is
important to note that these two processes apply on the triplet as a whole (i.e. Gene/RNA, protein and
modified protein nodes) and not to a partial set of nodes within the triplet.
1.8 Modeling mutations
Like triplet duplications and deletions, mutations happen at any time point and in any organism.
Mutations always target one or more parameters in a particular node. During a mutational event, one of
the following node-specific parameter groups is altered:
•
•
•
•
Target node midpoint (30 ) and slope (40 ).
Interaction weights (/01), regulator specific midpoint (301 ) and slope (4,01 ).
Basal expression parameter (basali).
Degradation parameter (OPQR0 )
The parameters within each group were specifically selected in order to match their biological
counterparts. A mutation in the first group captures a sequence or conformational change in the target
node that has an effect in its regulation function as a whole. A mutation in the second group models the
case where a change in residue(s) or base pair(s) – either in the regulator or target node – results in a
change of the interaction affinity between the target and the regulator node. This type of mutation is
regulator-target specific since it does not affect other regulatory interactions associated with the target
node. Finally a mutation may change the basal expression parameter (e.g. promoter becomes more/less
leaky) or the degradation parameter (e.g. point mutations in RNA or protein that affects its stability).
In addition to the type of parameters that a mutation targets, mutations can also be categorized based on
the severity of the change. As such, mutations can be grouped in two classes: Mild and Strong mutational
events. In mild mutational events, the parameters drift according to a Gaussian distribution around their
current value, whereas in strong mutational events they can take on any random value within the valid
(initial) range. Strong mutations are relatively rare events in the course of the experiment and they model
large impact changes that can result in loss – or even reversal – of function (e.g. an activating
transcription factor that becomes a repressor for a specific target gene). Figure S3 demonstrates the
variety of possible mutational events accompanying the evolution of a toy oscillator circuit.
1.9 Biochemical parameters and organism attributes
Organism parameters can be classified into three categories and are reported here with the ranges we
used during initialization in low mutation rate (LM) experiments. Parameters are sampled from a uniform
distribution, unless otherwise stated.
•
General parameters:
o Energy: The energy that the organism has at each time point (initially 8 · 10V energy
units).
o Fitness: The Pearson-correlation between the time series vectors that represent the
pattern of response pathway expression and abundance of environmental resources.
o Maintenance Energy Cost: Energy cost per unit time associated with the maintenance of
nodes. The maintenance energy cost for node i has a value of Costi = xi∙M energy units,
where xi is the node-specific maintenance cost and M the mutation modifier parameter.
For simplicity, in our experiments xi = 1 for all nodes. The mutation modifier per unit time
models the cost of DNA repair and other proofreading mechanisms and is the relative ratio
between the mutation rates at the beginning of the experiment to its current value. Thus,
an organism evolved to have lower mutation rate has a higher mutation modifier
parameter M (here M is equal to one since mutation rate per organism is constant).
o Response pathway cost: Energy cost per unit time associated with maintaining the
response pathway (modified protein RP1) active. It is linearly proportional to the number
of RP1 molecules present in the organism and it is given by costresponse_pathway = z∙P, where z
is the default cost for one protein per time unit (here z = 10) and P is the number of RP1
molecules.
•
•
Mutation Specific parameters:
o Triplet Creation Probability: The probability of triplet creation in each time unit (10-7 in
LM experiments).
o Triplet Destruction Probability: The probability of triplet destruction in each time unit
(10-7 in LM experiments).
o Mild Mutation Probability: The probability of mild mutation in any node per unit time(10-6
in LM experiments).
o Strong Mutation Probability: The probability of strong mutation in any node per unit time
(2∙10-7 in LM experiments).
o Evolvability: The rate of change in the creation, destruction, strong and mild mutation
probabilities. This is used only in experiments that allow organisms to vary their mutation
probabilities during the course of evolution (here 0.2 in all experiments).
Kinetic and Network related parameters:
o Number of Nodes: The total number of nodes in the cell (initially from 0 to 30).
o Node abundance: Abundance of molecules corresponding to a node (initially from 0 to 30).
o Connectivity: The fraction of actual to all possible connections in the internal network
(initially from 0.01 to 0.20).
o Basal: A vector that contains the basal levels of transcription, translation and modification
for all nodes in the internal network (initially from 0 to 0.5).
o Degradation: A vector that contains the degradation rates of all node values in the internal
network (initially from 0.05 to 1).
o Interaction matrix: A matrix with all regulatory dependencies between nodes and
environmental signals. Distribution of interaction weights initially obeys a power law with
W 1.5 while the connection probability of an environmental signal to a node decreases
exponentially with the number of nodes (in the same organism) it is already connected to.
o Target Slope and Midpoint: Two vectors that contain the slope and midpoint of the
primary target-specific sigmoid function that is used to model the regulation of a specific
node by other nodes and signals (initially slope ranges within [1,3] and midpoint within [1,1]).
o Regulator Slope and Midpoint: Two vectors that contain the slope and midpoint of a
secondary regulator-specific sigmoid function that is used to model the effect of regulation
by a node or signal of another node (initially the regulator slope ranges in [1,4] and the
regulator midpoint in [0,10].
1.10 Modeling organism-environment interactions
Organisms compete in an environment of constant population size that may contain any number of
signals with arbitrary or predefined characteristics. Various events that have a deleterious or beneficial
fitness effect on the organisms occur during the time course of an experiment. The temporal pattern of
event occurrence is flexible, ranging from purely stochastic to perfectly periodic. In the present work
experiments have a single event type, namely the presence of energy resources that organisms can
harvest through the expression of a metabolic pathway that is represented by a modified protein called
“response protein 1” (RP1). Environmental signals can be programmed to correlate – partially or fully with the future occurrence of an event, making it possible for organisms to predict environmental
trajectory. To achieve this, the organisms have to “learn” how to extract environmental information from
one or more signals. A specific node can couple/decouple its expression to an external signal as a
consequence of random mutations. Signal values can range from -1 to 1 and their regulatory effect is
directly proportional to that value.
The creation and destruction of any RNA, Protein or Modified Protein has an energy cost. Of special
interest is RP1, the modified protein of triplet 1, whose creation cost is ten times more than that of other
proteins. In addition to its high creation cost, the response protein has a maintenance cost per unit time
that is proportional to the number of its molecules present at any one time. This models well the real
metabolic machinery in cells, whose expression and maintenance is energetically costly. The probability
of death decreases exponentially with the energy level of the organism and approaches unity as energy
levels get close to zero. In order to keep the population size constant, a randomly generated organism is
placed in the population after the death of another.
The amount of energy that organisms acquire from the environment per unit time is proportional both to
the amplitude of the resource and to the response pathway expression. When an organism’s energy level
reaches a predefined threshold (in our analysis it is set to twice the initial energy level), the organism
undergoes mitosis. During mitosis, the organism is cloned and its energy level is reduced to half. The
newly created daughter cell takes the place of the organism with the least energy.
1.11 EVE simulation algorithm
We developed two different algorithms that simulate our evolutionary environment and obtained similar
results with both. Although the results presented here are derived from the real-time version of the
simulator, both generation-based and real-time implementations are described for the sake of
completeness.
1.11.1 Generation based simulator
At the beginning of each experiment, a fixed-sized population of organisms that receive a predefined
amount of energy units is created. During a given time duration, organisms undergo mutations while their
node values are continuously updated. At the end of the time interval, a new generation of organisms is
formed by “sampling with replacement” from the current population. The probability for an organism to
be selected for the next generation is directly proportional to the energy that it has at that point. Each
time a new generation is formed, the energy of all organisms is initialized. The simulation ends after a
predefined number of generations.
1.11.2 Real-time simulator
At the beginning of each experiment, a fixed-sized population of organisms that receive a predefined
amount of energy units is created. At each time point during the experiment, organisms can mutate,
divide and die in a parallel and asynchronous manner as described in the previous section. Statistics are
gathered at periodic time intervals (epochs) without this affecting the evolutionary trajectory of the
experiment. The experiment ends after a predefined amount of time units have elapsed.
1.12 Auxiliary algorithms
1.12.1 Deconstruction
The deconstruction algorithm was developed to assess the link/node significance and derive the minimal
internal network (i.e. the core computational module) of an organism. The algorithm deconstructs
networks in either a “parallel” or “serial” mode.
In parallel deconstruction, the significance of each individual link in the internal network of an organism
is evaluated within a time window which is at least an order of magnitude larger than the interval
between two epochs. During this evaluation process, a link gets knocked out and the fitness of the
resulting organism is assessed. Regardless of the link significance, each link is restored in the internal
network at the end of every evaluation. The program terminates once the fitness effect of all links is
assessed.
Serial deconstruction operates similarly with the difference that if a link proves to be insignificant (less
than 5% fitness change with respect to the initial fitness), it is not restored in the original network, which
leads to systematic pruning of non-essential links. To cope with redundant pathways in the network,
serial deconstruction performs multiple randomizations (permutations) of link assessment.
While parallel deconstruction is able to find the fitness impact of a single mutation in the network, serial
deconstruction enables us to recover redundancies and minimal networks responsible for a particular
phenotype. Each link significance is assessed multiple times (3 in our analysis) both in parallel and serial
deconstruction, while serial deconstruction creates several (3 in our analysis) permutations of the order
by which links get knocked out in order to uncover coexisting minimal networks. The results of parallel
and serial deconstruction can be visualized as a fitness matrix (eg. Fig. S13-A), depicting the negative
fitness fraction change between the initial and mutated phenotypes.
1.12.2 Epistasis analysis
Epistasis is defined as any non-additive interaction between two or more distinct mutations, such that
their combined effect on a phenotype deviates from the sum of their individual effects. Here, we compare
the fitness effect of link deletions in an organism’s internal network, both when two deletions occur
independently or concurrently as a pair. We use a variation of the parallel deconstruction algorithm that
knocks out two links simultaneously at each evaluation cycle in order to assess the epistatic potential of
link pairs, as manifested in their combined fitness effect (Fig. S13C, S14A).
1.12.3 Reverse engineering
As an alternative approach to network inference, we used the ARACNE algorithm (S8) which is based on
mutual information analysis of expression data. We conducted time series experiments and gathered the
expression signatures of all nodes in any given network. To accurately mimic “microarray” experiments,
we collected node expression data under different environmental conditions and stimuli, as for example
forcing one of the signals up or down. Figure S15B shows the topology of the reconstructed network that
belongs to the organism in Fig. 4, where the reconstruction algorithm discovered 4 out of the 11 links that
are essential for the organism to retain its phenotypic behavior. However the reconstructed organism has
none of the phenotypic properties of its original counterpart, thus rendering it essentially unfit under the
original selection environment. Nevertheless, systematic analysis of reconstructed networks evolved
under various mutation rates suggests that network reconstruction can indeed be helpful in identifying a
portion of the essential links in the biochemical network, possibly at the expense of a high false-positive
rate (Figure S16).
1.13 Additional simulations
All experiments presented here have a fixed population size of Np= 200 and time duration D = 1.8∙107
time units. Relative mutation rates are 1:1:10:2 for creation, destruction, mild and strong mutation
probabilities respectively. To obtain statistical significance for each environment and mutation rate
combination, we ran 64 independent simulations. Based on the selection pressure in the environment,
experiments can be classified into the following categories:
•
•
Delayed Gates: Signals and resource are related by OR, AND, XOR, NAND, NOR dynamic logic
functions.
Multi-gates: Signals and resource are interchangeably related by combinations of OR, AND, XOR,
NAND, NOR dynamic logic functions.
•
•
•
Oscillators: Selection pressure to evolve oscillatory expression of RP1 with or without a periodic
guiding signal.
Bi-stable Switches: Selection pressure to evolve bi-stability in environments where two
environmental signals operate as ON/OFF pulse switches.
Duration/variance Locking: Selection pressure to evolve networks that predict the duration of an
environmental resource that has fluctuating duration or phase variance.
Figure S4 summarizes the correlation between the environmental signals and the resource presence in
each experiment type. Figures S9-S11 depict minimal networks and/or phenotypes of representative
organisms that belong in the above categories (evolved oscillatory and multi-gate organisms have large
complex networks and thus are not included here).
2. Experimental Methods
2.1 Bacterial strains and growth conditions
E. coli strain MG1655 and its derivates were used in all physiology and evolution experiments. A library
of Tn5 transposon mutants (S9) with a diversity of ~106 was used as the starting point for evolution
experiments in order to: 1) increase the diversity of genetic perturbations on which selection acts and 2)
allow marking and monitoring of individual mutants within the evolving population. Bacteria were
grown in batch or bioreactor vessels in M9 minimal media supplemented with 0.4% glucose.
2.2 Controlled temperature and oxygen perturbations
Physiological perturbations were carried out under a controlled environment in the context of bioreactor
(Bioflo 110, New Brunswick Scientific) growth. Thermoelectric sensors and heaters were used to shift
temperature profiles between 250 C and 370 C, and polarographic dissolved oxygen sensors (Mettler
Toledo) and nitrogen gas was used to rapidly change oxygen saturation between anaerobic (0% dissolved
oxygen) and aerobic (16-21% dissolved oxygen) condition. The cultures were maintained in exponential
phase (O.D.600 0.2-0.4) through controlled dilution, where fresh media is pumped in and spent media is
pumped out at a controlled rate. Prior to the perturbations, cells were maintained in the pre-transition
environment for at least eight generations (8-24 hours). All experiments were performed in duplicate,
with high reproducibility (Fig. S17). Unless specifically indicated, all perturbations were performed
during exponential growth, and cells were harvested at variable intervals and rapidly mixed in an ice-cold
ethanol/phenol solution (5% water-saturated phenol in ethanol) in order to stop mRNA degradation.
After spinning down the cells at 5000 rpm for 5 minutes at 40 C, the media was discarded, and cell pellets
were frozen on dry-ice/ethanol for storage at -800 C.
2.3 Microarray transcriptional profiling
A hot-phenol procedure was used to extract total RNA. Cell pellets were lysed with 500 µl TE (pH8.0), 50
µl 10% SDS and lysozyme (0.5 mg/ml). Total RNA was extracted sequentially with phenol/chloroform
(preheated to 640 C), followed by chloroform/isoamyl alcohol. RNA was precipitated with 1/10 volume
of 3M NaOAc (pH 5.2) and 2 volumes of ethanol. After incubating overnight at –200 C, samples were spun
down and pellets were washed with ice cold 70% ethanol (prepared with DEPC- H2O). RNA was
resuspended in water, DNase treated (RQ1 RNase-free DNase/ Promega, WI) and purified using an
RNeasy purification kit (Qiagen, CA). Fluorescent cDNA was synthesized from total RNA with 15 µg of
total RNA serving as template. Cyanine-labeled Cy3 or Cy5-dUTP (Amersham Bioscience) and pdN6
random hexamers (GE healthcare) were used in reverse transcription reactions utilizing SuperScript II
(Invitrogen, CA) for cDNA synthesis. Fluorescent genomic DNA was generated as described in (S9).
Genomic DNA served as a universal reference for all hybridizations. E. coli genomic DNA was fragmented
(500 – 2000 bp) using mechanical shearing, and subjected to Cy3 or Cy5-dUTP labeling using the
BioPrime DNA labeling system (Invitrogen, CA). Following purifications through CyScribe GFX
Purification Kit (Amersham Biosciences), Cy-dye labeled genomic DNA and RNA-derived cDNA probes
were combined with a hybridization buffer (5X SSC, 0.1% SDS, 50% formamide, and 10 μg Salmon sperm
DNA) and hybridized for ~ 16 hours at 420 C to a DNA microarray containing 95% of all ORFs in the E. coli
MG1655 genome (S9). All microarray experiments were carried out in biological duplicates and showed
high reproducibility (Fig. S18).
2.4 Microarray data analysis
Microarray slides were scanned on a GenePix 4000B scanner (Axon Instruments), and fluorescence data
for each of the duplicates on each slide were analyzed using GENEPIX PRO 5.0 software. In a quality-
control step, elements with poor spot morphology or exhibiting uneven hybridization caused by dust
particles or scratches, were flagged manually and excluded from further analyses. After local background
subtraction and global normalization relative to the genomic DNA reference, duplicate measurements on
the same array were averaged, yielding a single vector for each hybridization. Seven time-points were
assayed for each perturbation, corresponding to 0, 4, 8, 12, 20, 28, and 44 minutes post transition. After
relative scaling, biological duplicates from all experiments were averaged, leading to a single set of timeseries data for each perturbation. To compare our results to other perturbations, we combined our
temperature/oxygen transcriptional responses with previously published studies of UV (S10) and
osmolarity stress (S11) in E. coli. For most of the analyses, we focused on changes in gene expression that
occurred at various points (e.g. 20 minutes) relative to the zero time-point reference. Log (base 2)
transformations of fold-changes were utilized in determining the global correlation between various
perturbations (Fig. 5B, 6D). Sets of differentially expressed genes were defined as showing 1.5 fold or
higher increase/decrease in gene expression relative to the time zero reference. Variation of this
threshold from 1.33 to 2.00 gave very similar results. The hyper-geometric distribution was used to
determine the chance-probability of observing overlaps in sets of differentially regulated genes (Fig. 5A).
The combined expression dataset (consisting of 2612 genes across 34 conditions/time-points) is
available and can be downloaded from our web site.
2.5 Experimental evolution under an ecologically incoherent dynamic environment
A diverse library of Tn5 transposon mutants (S9) was used as the starting point for evolution of E. coli
MG1655 for optimal growth under a dynamic environment where temperature and oxygen transitions
were positively correlated with a 40 minute time-lag (Fig. 6A). In order to avoid selection for periodic
behavior, intervals of constant temperature/oxygen were sampled from a Gaussian distribution with a
mean of 150 minutes (37C/O2+) and 300 minutes (25C/O2-) and standard deviations of 30 minutes
(37C/O2+) and 60 minutes (25C/O2-), respectively. In total, 44 cycles of selection were carried out over
the course of 384 hours. The starting population contained ~1x1010 transposon mutants (diversity of
5x105) growing in exponential phase (maintained between O.D.600 0.20-0.30 by continuous pumping of
fresh media and discarding spent media) in a volume of ~500 ml of M9 media supplemented with 0.4%
glucose. Periodic sampling of the culture allowed us to monitor growth-rate and archive representative
samples of the population at -800 C. The presence of a single randomly inserted transposon in every
genome allowed us to monitor population diversity and track high-fitness mutants during the evolution
experiment through the application of a gel-based genetic footprinting procedure (Fig. S20). The
amplification of transposon insertion sites from a sample of 100 colonies isolated at the end of selection
identified a highly abundant mutant that made up 55% of the population. This “evolved” mutant,
designated as aldB::Tn5 was used in subsequent fitness evaluation and physiology experiments.
2.6 Competition between parental and evolved strains
To establish the fitness advantage of the evolved strain, we carried out competition experiments within
the setting of the dynamic environment imposed during the evolution. The evolved and the parental
strains were distinguishable through the differential expression of a LacZ marker. This system allowed us
to easily quantify the relative fitness advantage of the evolved mutant using colony counting on plates
(S12). A 60-hour competition experiment established the strong fitness advantage of the evolved strain
(Fig. S21).
3. Figures and tables
0 16.19 0
[0
0
0
Z
0
0
0
W =Z
0
0
Z0
Z0 9.91 0
Y0
0
0
0
0
13.64
0
0
0 b
a
0
0
0 a
0 3.77
0 a
0
0
0 à
0
0
0
Figure S1 | A randomly created internal network and corresponding interaction matrix. Red and
blue edges denote activation and repression respectively. The absolute value in each edge corresponds to
the weight of regulation while a positive/negative sign represents activation/inhibition respectively.
Green edges depict creation in the case of mRNA transcription and protein translation, or transformation
in the case of protein modification.
A
fi1(x1,m
˜ i1,s˜i1)
fi2(x2,m
˜ i2,s˜i2)
Gi
...
Pi
fin(xn,m
˜ in,s˜in)
B
C
0.8
Probability
1.0
˜ = 2.5
m
˜ = 3.75
m
m
˜ =5
m
˜ = 6.25
m
˜ = 7.5
0.6
0.4
s˜ = 1
s˜ = 2
s˜ = 3
s˜ = 4
0.6
0.4
0.2
0.2
0
˜ =5
m
0.8
Probability
1.0
s˜ = 2.5
1
2
3
4
5
6
7
8
Number of Molecules
9
10
0
1
2
3
4
5
6
7
8
9
10
Number of Molecules
Figure S2 | Expression model and molecule production probability. (A) A two level sigmoid function
is used in the expression model. The molecule production probability is represented as a second level
sigmoid that has as input the sum of all individual regulatory effects. (B) Plot of function , , ̃ "
for different values of the midpoint, while keeping the slope constant , +* 2.5 (C) Plot of function
, , ̃ " for different values of the slope, while keeping the midpoint constant (
5
Figure S3 | Impact of mutational events on network parameters. (A) Network topology of a toy
oscillator, that was evolved based on an initially user defined, non-functional design. Proteins MP0 and P0
activate (red lines) G1, while MP1 represses (blue line) G0 and G1. The P0-G1 activation (interaction
weight c ) is depicted with a dashed line since it is deleted during the evolution. (B) Expression profile
of both triplets over time. In each plot, blue, green and red waveforms correspond to RNA, protein and
modified protein levels respectively. (C) Key parameters and their values in the initial state, after type1,
type2 and type3/4 mutations. The expression profile of both triplets for each state is shown at the bottom
of the figure.
Figure S4 | Correlation of environmental signals with resource abundance for various
environments. The resource abundance (red) during a specific signal pattern (black) is given for all
resource-signal correlations mentioned in the text. Depending on the selection pressure there can be
none, one or more signals present in the environment: oscillators have zero or one, toggle switches and
gates have two signals while multi-gates have three. In a toggle switch, Signals 1 and 2 serve as ON and
OFF switches respectively. In a multi-gate the third signal indicates a switching event, for example a
transition from an AND to an OR logic.
Figure S5 | Fitness (Pearson-correlation) Histogram for all delayed logic gate experiments. Each
subplot depicts the histogram of Pearson-correlation between response protein expression and resource
abundance of the fittest organism in the final populations (64 final populations per mutation
rate/environment) of Supplementary Table 1. Each row corresponds to a gate type (AND, OR, XOR, NOR,
NAND) and each column to a mutation rate (VLM, LM, MM, HM). For each subplot, x-axis and y-axis
represent Pearson Correlation and number of experiments respectively.
Signal S2
Signal S2 Signal S1
Signal S1
A
B
Response
Resource
Signal S1
Signal S2
Response
Resource
Signal S1
Signal S2
Time
1.14
MP0
P0
1.09
2.60
RP1
MP2
P1
MP3
P2
MP0
2.03
P3
12.28
RP1
P0
Promoter 0 G0
-0.63
RNA1
1.27
RNA2
4.21
Promoter 1 G1
P1
1.58
1.83
RNA0
Time
3.39
-1.34
1.86
RNA3
Promoter 2 G2
Promoter 0 G0
1.53
P2
2.34
1.86
1.85
RNA2
Promoter 1 G1
Signal S2 Signal S1
C
-1.24
-1.27
RNA1 1.11
RNA0
G3 Promoter 3
MP2
1.13 1.23
Promoter 2 G2
Signal S2 Signal S1
Response
Resource
Signal S1
Signal S2
D
Response
Resource
Signal S1
Signal S2
Time
Time
1.05
1.37
0.97
1.86
-7.74 -7.74
1.37
MP0 9.70
RP1
MP2
MP3
-5.16
-8.82
5.09
-8.82
-8.82
5.90
-8.82
-7.86
4.28
-6.39 P1
0.97 1.37 P2
P0
P3
1.37
1.91
-1.841.91
2.58
2.30
2.13
2.04
1.85
-1.63
1.85
5.46
RNA0
RNA1
RNA2
RNA3
2.36
2.60
Promoter 0
G0
Promoter 1
MP0
RP1
3.60
-8.26
-10.79
-1.38
-9.74
7.25
P0
2.26
-5.41 7.38
2.71
RNA0
0.37
2.60
G1
Promoter 2
G2
G3 Promoter 3
G0
Signal 2
1.23
Promoter 0
7.35
P1
8.71
-8.87
3.84
-2.26
RNA1
Promoter1
G1
Signal 1
E
Response
Resource
Signal S1
Signal S2
Time
-11.87
1.93 0.33
0.36 1.90
-7.69
MP0
RP1
MP2
1.76
3.07
-9.86
-8.80
7.35
P0
-10.03
-11.56
4.25
28.82
3.23
RNA0
4.62
P1
-7.70
G0
-6.65
11.86 2.72
0.90
7.41
6.97
RNA1
Promoter 1
RNA2
1.20
-0.62
Promoter 0
P2
-7.19
G1
G2
Promoter 2
Figure S6 | Network topology of the fittest cell during the evolution of a delayed XOR. The networks
portrayed here correspond to the fittest cell at epochs (1-5) of [Fig. 3]. Activation and inhibition are
represented by red and blue arrows respectively, while solid and dashed lines indicate essential and nonessential links. The phenotypic behavior of each organism is depicted in the upper right corner. (A, B) The
initially random, low fitness network evolves to a network that partially infers the resource fluctuation in
the environment by coupling, although poorly, to signal S1. (C) Later, organisms evolve to predict all
resource peaks, although their behavior is noisy and not matched well with resource abundance. (D)
Once a stable phenotype emerges, it takes over the population despite the fact that organisms with higher
fitness were present in the population. (E) A triplet duplication (triplet 0 to triplet 2) and subsequent
optimizing mutations give rise to the final fit organism.
Fitness
Figure S7 | Fitness trajectory in a delayed XOR experiment. The evolution of a separate XOR
experiment under the same conditions as that of [Fig. 3] is depicted. (A) Fitness trajectory of a delayed
XOR experiment where the presence of any of the two signals alone is necessary and sufficient condition
for future resource abundance. Red and black lines correspond to the highest and mean fitness in the
population at each epoch respectively. Fitness is defined as the Pearson-correlation between resource
abundance and response protein expression. (B) The phenotypic behavior of the fittest organism at
different points along the evolutionary trajectory. Each subplot consists of four rows: the first row
depicts the abundance profiles of the RNA (blue), protein (green), and modified protein (red) of the
response pathway. The second, third and fourth rows correspond to the resource abundance, and
environmental signals S1 and S2 respectively.
Figure S8 | Network topology of the fittest cell during the evolution of a delayed XOR. The networks
portrayed correspond to the fittest cell at epochs (1-4) of Fig. S7. Activation and inhibition is represented
by red and blue arrows, while solid and dashed lines indicate essential and non-essential links
respectively. The phenotypic behavior of these organisms is depicted in the upper right corner. The
initially random and unfit network progressively evolves to a network that can accurately predict the
resource fluctuation in the environment. (A) Initial random network. (B) Partial correlation evolved
within the first 500 epochs. (C) A duplication event of triplet 0 facilitates the evolution of a higher fitness
organism. (D) Further optimization by less severe mutations and emergence of the final phenotype.
Signal S2
Signal S1
A
Signal S1
Signal S2
B
Response
Response
Resource
Resource
Signal S1
Signal S1
Signal S2
Signal S2
Time
Time
13.89
MP3
RP1
P3
P1
MP4
MP6
RP1
P4
P6
P1
RNA4
RNA6
2. 73
3.03
11.66
RNA3
RNA1
RNA 3
RNA1
Promoter3
G3
Promoter1
G1
Promoter4
G4
G6
Signal S1 Signal S2
C
Promoter 3
Promoter 6
G3
RNA 4
-1. 04
-1. 21
Promoter 1
G1
G4
Signal S2
Signal S1
D
C
Promoter 4
Response
Resource
Signal S1
Signal S2
Response
Time
Resource
Signal S1
MP5
Signal S2
MP0
Time
-5.83
P5
-1.99
P0
-1.21
1.10
-1.22
RNA5
RNA0
1.34
MP4
13.42
Promoter 5
G5
P4
MP 0
Promoter 0
G0
MP3
RNA4
MP2
RNA1
P1
Promoter 4
RNA0
G0
2.83
2.41
7.39
Promoter 0
P1
2.11
RP1
0.99
P0
RP1
-4.05
RNA 1
G1
G4
P3
-5.52
RNA 3
Promoter1
Promoter 3
P2
-1.47
Promoter 1
G1
RNA 2
G3
Promoter 2
G2
Figure S9 | Minimal networks of delayed dynamic logic gates evolved in environments with low
mutation rate. (A) Delayed OR gate where the response pathway is activated through protein P3,
modified protein MP6 or both simultaneously. The presence of either signal S1 or S2 is sufficient for the
activation of the response pathway. Protein P4 catalyzes the modification of P1 to MP1 and increases the
overall fitness of the system (Pearson correlation between MP1 expression and resource presence
increases to 0.82 from 0.63) (B) Delayed NOR gate where the absence of both signals is necessary for the
activation of the response pathway. (C) Delayed AND gate where the response protein is expressed only
when both signals are present. Signal 2 initiates the transcription of G1, while high MP0 concentration is
necessary for the translation of P1. The latter happens only in the presence of signal S1, since it regulates
the expression of G0. (D) Delayed NAND gate where the response protein is not expressed when both
signals are present simultaneously.
Figure S10 | Minimal network and behavior of an ON/OFF switch. (A) Signal S1 (ON signal) activates
the transcription of G1 that leads to the production of RNA1, P1 and RP1. RP1 acts in a positive feedback
loop that activates the translation of P1 and locks the creation of RP1 molecules in the absence of Signal
S2 (OFF signal). Once a Signal S2 pulse is present, MP7 and P7 directly or indirectly (through inhibition of
RNA3, a positive regulator of P1-MP1 modification) inhibit the expression of MP1. (B) Signal S1 and
Signal S2 pulses indicate the start and the end of resource presence in the environment respectively.
Response protein expression (first row, red line) correlates well (>0.9 Pearson Correlation) with resource
abundance (second row).
Number of Occurencies
Figure S11 | Evolution of a duration inference mechanism. (A) In an experiment where the duration
of the resource presence is variable, organisms have evolved to infer duration variability and highly
correlate the expression of response protein with the presence of resource. The duration inference
mechanism is based on node coupling to a continuous environmental signal (Signal S3) whose phase is
indicative of the environmental variance. (B) Distribution over 90 epochs of response protein duration
when the duration of resource presence in the environment interchanges between 4000 ± 300 and 6000
± 300 time units. Response protein expression follows a bimodal distribution with two peaks that
correspond to short (red) and long (blue) expression periods. The duration of a response pulse is defined
as the time interval between the ascending and descending edge, measured at the 50% of peak expression
amplitude.
Number of Cells
Number of Cells
Figure S12 | Fitness and network-size trajectory during the evolution of the delayed XOR network
([Fig. 3]). Each point along the time axis corresponds to an epoch, i.e. 4,500 time units in the experiment.
(A) A time series depiction of population fitness for all cells (organisms). There are two fitness fixation
points at 0.6 and 0.95 that occur around 1200 and 3000 respectively and correspond to the emergence of
the one and two signal “mutual exclusive” computational modules, as described in the paper. After the
second dominant fixation we observe a decline in the variation of population fitness. (B) Node number
distribution in the internal non-minimal network of all cells. Nodes are always created and destroyed in
triplets (gene, protein and modified protein nodes), hence the discrete nature of the plot. Although the
node number variance decreases after the first fixation (where only six nodes comprise the minimal
network), there is still triplet duplication events that ultimately lead to the optimal phenotype (which
requires a minimum of nine nodes).
MP2
P2
RNA 2
RP1
P1
RNA 1
MP0
P0
RNA 0
MP2
P2
RP1
RNA 2
P1
RNA 1
MP0
P0
Connection ID
RNA0
Figure S13 | Network deconstruction in the delayed XOR experiment ([Fig. 3]). (A) Fitness effect of
single link knockouts (parallel deconstruction) illustrates link essentiality in the internal network. The
fitness impact of links vary from zero (green) to one (dark red) and corresponds to the fitness fraction
change that the link severance causes with respect to the fitness of the original network (a 100% fitness
change is denoted by 1, 42% by 0.42 and so forth). Matrix elements that depict non-existent links are
colored blue. (B) Knockouts of node pairs reveal functional similarities between nodes: simultaneous
deletion of all links in nodes P0 and P2 reveals strong negative epistasis. (C) Epistatic interaction matrix
of all 81x81 possible link pair knockouts. Each matrix element shows the fitness effect of a specific link
pair knockout. Non-existent link pairs are colored blue.
A
Pair Link ID
(0,3) (0,5) (1,4) (1,5) (2,0) (2,6) (3,0) (3,1) (3,3) (3,5) (3,6) (3,7) (4,0) (4,3) (4,4) (4,6) (5,1) (5,2) (5,5) (5,7) (5,8) (6,3) (6,5) (7,4) (7,5) (8,0) (8,6)
(0,3)
0.34
0.96
0.34
0.44
0.02 0.46
(0,5)
0.34
0.95
0.33
0.46
0.02 0.46
(1,4)
0.34
0.96
0.34
0.42
0.02 0.44
(1,5)
0.35
0.95
0.33
0.44
0.45
(2,0)
0.34
0.95
0.33
0.44
0.03 0.48
(2,6)
0.34
0.95
0.34
0.46
0.02 0.44
(3,0)
0.34
0.95
0.34
0.43
0.02 0.46
1
0.9
0.8
(3,1) 0.34 0.34 0.34 0.35 0.34 0.34 0.34 0.33 0.34 0.94 0.34 1.00 0.34 0.34 0.33 0.35 0.34 0.34 0.35 0.98 0.34 0.34 0.34 0.34 0.34 0.34 0.33
(3,3)
0.34
0.97
0.34
0.45
0.7
0.02 0.47
Pair Link ID
(3,5) 0.96 0.95 0.96 0.95 0.95 0.95 0.95 0.94 0.97 0.96 0.95 0.94 0.96 0.96 0.95 0.95 0.99 0.95 0.91 0.98 0.95 0.95 0.95 0.96 0.96 0.96 0.95
(3,6)
0.34
0.95
0.34
0.41
0.02 0.44
0.6
(3,7) 0.34 0.33 0.34 0.33 0.33 0.34 0.34 1.00 0.34 0.94 0.34 0.34 0.34 0.34 0.34 0.34 1.00 0.34 0.35 0.34 0.33 0.34 0.33 0.33 0.34 0.34 0.33
(4,0)
0.34
0.96
0.34
(4,3)
0.34
0.96
0.34
0.33
0.95
0.34
0.35
0.95
0.34
(4,4)
0.02
(4,6)
0.20
0.20
0.46
0.02 0.42
0.47
0.02 0.53
0.43
0.46
0.44
0.46
0.5
0.02
(5,1) 0.44 0.46 0.42 0.44 0.44 0.46 0.43 0.34 0.45 0.99 0.41 1.00 0.46 0.47 0.43 0.44 0.44 0.42 0.49 0.99 0.44 0.47 0.41 0.42 0.44 0.44 0.43
(5,2)
0.34
(5,5)
0.95
0.34
0.42
0.03 0.02 0.02 0.35 0.02 0.91 0.02 0.35 0.02 0.02
0.4
0.46
0.02 0.47 0.02 0.02
0.49
0.02 0.02 0.03 0.02
0.3
(5,7) 0.46 0.46 0.44 0.45 0.48 0.44 0.46 0.98 0.47 0.98 0.44 0.34 0.42 0.53 0.46 0.46 0.99 0.46 0.47 0.44 0.43 0.47 0.47 0.48 0.42 0.46 0.48
(5,8)
0.34
0.95
0.33
(6,3)
0.34
0.95
0.34
(6,5)
0.34
0.95
0.33
(7,4)
0.34
0.96
0.33
0.42
0.02 0.48
(7,5)
0.34
0.96
0.34
0.44
0.02 0.42
(8,0)
0.34
0.96
0.34
0.44
0.03 0.46
(8,6)
0.33
0.95
0.33
0.43
0.02 0.48
0.02
0.44
0.02 0.43
0.47
0.02 0.47
0.02 0.41
0.47
B
0.2
0.1
0
C3
C2
C1
P1
P1
RNA 1
P1
P2
G1
P0
G1
RP1
G1
P2
P1
28.82
RNA 1
RP1
P0
RP 1
RP 1
11.86
P0
7.41 6. 97 P2
Promoter 1
G1
P0
- 8. 80
-9. 86
P1
P2
-15.16
RNA 1
Promoter 1
G1
Figure S14 | Epistatic Analysis of a delayed XOR network ([Fig. 4]). (A) Epistatic matrix of existent
pair link knockouts derived from the augmented epistatic matrix of Fig. S13c. Only fitness changes more
than 1% are shown. Each link pair is represented by the target-regulated and its source-regulator node
respectively. Nodes are numbered from 0 to 8 from the lower triplet index to the higher and from RNA to
Protein to Modified Protein. For example, matrix row (3,1) denotes the knockout of the direct regulation
of G1 by P0. Strong negative epistasis is observed at the severance of both positive regulatory links from
P0 and P2 to G1 (link pair (3,1) and (3,7)), both inhibitory interactions between P0, P2 and P1
modification (link pair (5,1) and (5,7)), as well as their crosslink combinations: knockout of P0 (P2)
mediated activation of G1 and repression of P1 modification by P2 (P0) – which corresponds to link pairs
(3,1)-(5,7) and (3,7)-(5,1) respectively. Positive epistasis occurs during simultaneous deletion of both
inhibitory RP1 to G1 and self-activating RP1 to RP1 interactions, during deletion of repression/activation
pair P1-RP1 and P1-G1 (link pair (5,1) and (3,1)), and during deletion of P2-RP1 and P2-G1 (link pair
(5,7) and (3,7)). (B) Monochromatic clustering of epistatic interactions (Segre,D. et al. Nat. Genet. 37, 7783) reveals the existence of remarkably compact sub-modules that are responsible for keeping P1
expression on after a S1/S2 signal pulse (c1), response pathway activation (c2) and delayed behavior of
the circuit (c3).
A
G0
50
P0
45
MP0
40
G1
35
P1
30
RP1
25
G2
P2
20
MP2
15
S1
10
S2
5
R
0
0
1000
2000 3000
4000
5000
6000 7000 8000 9000
Time
B
S1
S2
MP0
RP1
P0
P1
G0
G1
MP2
P2
G2
E
+
+
-
NE RC
+
+
+
+
Figure S15 | Microarray analysis and reverse engineering. (A) Depiction of node expression (G0MP2), signals (S1-S2) and resource (R) abundance of the fittest organism evolved in a delayed XOR
environment ([Fig. 4]). The organism successfully predicts resource abundance and expresses RP1 to
harvest resources when present. (B) Network reconstruction derived from the application of an
information-theoretic reconstruction algorithm (ARACNE; Basso,K. et al. Nat. Genet. 37, 382-390) and
“microarray” data of network activity for various combinations of exposures to signal S1 and S2. Thick
and thin red lines represent essential (E) and non-essential (NE) links that were successfully
reconstructed (RC) respectively, while the blue dotted lines map to essential links that were not
recovered by the reconstruction algorithm. Black lines indicate links that are predicted in the
reconstructed network but which do not exist in the original network.
Sensitivity
Precision
60.00%
Original
40.00%
Minimal
20.00%
80.00%
60.00%
Original
40.00%
0.00%
ANDlm
ANDmm
ANDhm
ORlm
ORmm
ORhm
XORlm
XORmm
XORhm
NORlm
NORmm
NORhm
NANDlm
NANDmm
0.00%
1
3
5
7
9
11
13
Minimal
20.00%
ANDlm
ANDmm
ANDhm
ORlm
ORmm
ORhm
XORlm
XORmm
XORhm
NORlm
NORmm
NORhm
NANDlm
NANDmm
Precision
80.00%
Sensitivity
100.00%
100.00%
1
3
5
7
9
11
13
Figure S16 | Statistics of reconstructed networks using a mutual information-based algorithm
(ARACNE). We report the total number of links, link percentage, true positives (TP), false positives (FP),
de
de
dh
), sensitivity (
) and specificity (
) of
true negatives (TN), false negatives (FN), precision (
defge
defgh
dhfge
the reconstructed network compared both to the original and minimal network of each organism. The
two plots indicate the precision and sensitivity of reconstruction with respect to the original and minimal
network. Interestingly, the precision of reconstruction is higher for the original network while the
sensitivity of reconstruction is higher for the minimal network. The apparent asymmetry stems from the
fact that all links in the minimal network are essential and in that respect they have to be present in the
reconstructed network for it to fully exhibit its phenotype.
B
41.0
36.0
T
31.0
26.0
21.0
-20
0
20
40
60
80
Temperature (0 C)
Temperature (0 C)
A
D
20.0
15.0
O2
10.0
5.0
0.0
-20
0
20
40
60
Time after O2 shift (min)
80
31.0
T
26.0
21.0
-20
0
20
40
60
80
Time after temp.shift (min)
Dissolved O2 (%)
C
Dissolved O2 (%)
Time after temp.shift (min)
36.0
20.0
15.0
O2
10.0
5.0
0.0
-20
0
20
40
60
80
Time after O2 shift (min)
Figure S17. Temperature and oxygen perturbation profiles are highly reproducible. Temperature
and oxygen levels were precisely monitored and controlled using a New Brunswick BioFlo 110 bioreactor
(New Brunswick Scientific). All experiments were performed in duplicate with high reproducibility (red
vs. blue). (A) Temperature is shifted from 250 C to 370 C within a 20 minute interval while the culture is
maximally aerated (18% dissolved O2). (B) Temperature is shifted from 370 C to 250 C within a 20 minute
interval while the culture is maximally aerated (18% dissolved O2). (C) Oxygen saturation is downshifted within a 20 minute interval while temperature is maintained at 250 C. (D) Oxygen saturation is
up-shifted within a 20 minute interval while temperature is maintained at 370 C.
2
B
r = 0.969
10
10
10
1
0
-1
2
10
T up; 16 min. (2)
T up; 0 min. (2)
A
10
r = 0.951
1
10
0
10
-1
10
10
-1
0
1
10
10
10
2
10
10
0
1
10
10
2
T up; 16 min. (1)
T up; 0 min. (1)
2
10
T up; 16 min. (1)
C
-1
r = 0.808
1
10
0
10
-1
10
10
-1
10
0
1
10
10
2
T up; 0 min. (1)
Figure S18. Comparison of microarray expression profiles. High reproducibility of independent
experimental replicates: (A) temperature up perturbation, 0 minutes, (B) temperature up perturbation,
16 minutes. (C) Lower correlation reflecting differential expression between 0 and 16 minutes post
temperature up transition.
UP-REGULATED
DOWN-REGULATED
A
292
109
O2
B
316
292
89
292
O2
212
50
p < 10-31
51
296
373
p < 10-1
327
OSM
247
71
247
O2
p < 10-9
161
p < 10-28
T
43
O2
p < 10-11
230
T
O2
UV
78
247
O2
T
O2
D
p < 10-58
T
O2
C
198
354
p < 10-1
UV
32
335
p < 100
OSM
Figure S19. Transcriptional responses to ecologically correlated perturbations in temperature
and oxygen show statistically significant overlap. Sets of genes that are differentially expressed by a
factor of 1.5 fold show highly significant overlap when temperature and oxygen perturbations correspond
to ecologically correlated changes accompanying transitions between the outside environment and the
GI-tract (A, B). Significantly lower overlap is seen between the oxygen down-shift response and UV
response (C), and response to hyper-osmolarity stress (D).
B
A
transposon
chromosomal DNA
0
48
96 144 192 240 288
Lyse cells or isolate genomic DNA
from mutants
Digest with restriction
enzymes
Y-linker
Ligate to Y-linker
1st cycle of
PCR
PCR amplification of transposonadjacent genes
DNA
In vitro transcription
RNA
cDNA synthesis
cDNA
Figure S20. Parallel amplification of transposon-adjacent sequences and monitoring population
diversity. (A) Genomic DNA extracted from a population of transposon mutants exposed to a selection is
restriction digested, and the resulting fragments are ligated to a partially double-stranded Y-linker (Girgis
et al. PLoS Genetics 3(9): e154 (2007)). In the initial PCR cycle, primer annealing to a sequence within the
transposon and DNA polymerase mediated extension generates a double-stranded DNA template. In
subsequent PCR cycles, primers annealing to a region within the transposon and the top strand
complement of the Y-linker allows exponential amplification of transposon-adjacent sequences. An in
vitro transcription reaction using the PCR products generates RNA which, in the presence of reverse
transcriptase and modified nucleotides, generates fluorescently labeled cDNA. The fluorescently labeled
cDNA can then used in subsequent microarray hybridizations in order to get a genome-wide picture of
transposon insertions. (B) The DNA products resulting from the PCR amplification of transposon
adjacent sequences can be easily visualized on an ethidium-bromide gel, providing a qualitative picture of
population diversity throughout the evolution experiment. At the beginning, the population has maximal
diversity (smear at time zero). Distinct bands correspond to individual mutants that become increasingly
abundant as a function of selection.
white/red:
B
0 hr
12 hr
24 hr
50/57
86/71
392/99
Population fraction
A
36 hr
162/9
48 hr
60 hr
414/10
302/5
1
0.8
0.6
Evolved
0.4
Parental
0.2
0
0
12
24
36
48
60
Hours
Figure S21. Competition experiments between the parental and the evolved strain. (A) Differential
expression of a LacZ marker allows monitoring of relative growth in mixed competition experiments. (B)
The evolved strain rapidly outcompetes the parental over the course of 60 hours of growth under
ecologically incoherent correlations between temperature and oxygen.
VLM
LM
MM
HM
Total
AND
14/64
17/64
40/64
22/64
93/256
(36.3%)
OR
0/64
7/64
18/64
0/64
25/256
(9.8%)
XOR
0/64
4/64
13/64
0/64
17/256
(6.6%)
NOR
3/64
25/64
39/64
23/64
90/256
(35.2%)
NAND
0/64
4/64
5/64
11/64
20/256
(7.8%)
Total
17/320
57/320
115/320
56/320
(5.3%)
(17.8%)
(35.9%)
(17.5%)
Supplementary Table 1 | Success-rate of five delayed dynamic logic gate experiments across four
different mutation rates. All experiments were conducted for 4000 epochs (1.8∙107 time units). An
experiment is considered successful if the input-output characteristics of the fittest organism are
representative of the gate we select for. Operationally this translates to having the Pearson Correlation
between the resource abundance and the response protein expression higher than 0.7. From left to right
the columns represent environments that have very low mutation (VLM, triplet creation probability 10-8),
low mutation (LM, triplet creation probability 10-7), medium mutation (MM, triplet creation probability
10-6), high mutation (HM, triplet creation probability 10-5) rate. In each table entry the first value
corresponds to the number of successful experiments and the second to the total number of experiments
conducted. Other experiments with lower success rate included oscillators with (3/256) and without
(1/512) synchronization signal, bistable switches (2/64) and OR-XOR multi-gates (1/128).
References
S1. P. Francois and V. Hakim, Proc.Natl.Acad.Sci.U.S.A 101, 580-585 (2004).
S2. P. Francois, V. Hakim, E. D. Siggia, Mol.Syst.Biol. 3, 154 (2007).
S3. Rodrigo G., Carrera J., A. Jaramillo, CEJB 2, 233-253 (2007).
S4. O. S. Soyer, T. Pfeiffer, S. Bonhoeffer, J.Theor.Biol. 241, 223-232 (2006).
S5. A. Wagner, Proc.Natl.Acad.Sci.U.S.A 102, 11775-11780 (2005).
S6. K. A. Fichthorn and W. H. Weinberg, J.Chem.phys. 95, 1090-1096 (1991).
S7. D. T. Gillespie, J.Phys.Chem. 81, 2340-2361 (1977).
S8. K. Basso et al., Nat.Genet. 37, 382-390 (2005).
S9. H. S. Girgis, Y. Liu, W. S. Ryu, S. Tavazoie, PLoS.Genet. 3, e154 (2007).
S10. J. Courcelle, A. Khodursky, B. Peter, P. O. Brown, P. C. Hanawalt, Genetics 158, 41-64 (2001).
S11. K. J. Cheung, V. Badarinarayana, D. W. Selinger, D. Janse, G. M. Church, Genome Res. 13, 206-215
(2003).
S12. S. F. Elena and R. E. Lenski, Nat.Rev.Genet. 4, 457-469 (2003).